KNOD: In-Kernel Network Offload Device for GPU-Accelerated Packet Processing
summary
KNOD is a Linux kernel framework that natively compiles XDP BPF into GPU ISA and dispatches packet processing without ROCm or userspace, using a Radeon RX 6600 + ConnectX‑5 setup (~$450). Prototype reaches ~70 Mpps, with an XDP offload case showing compiles BPF to AMD GFX9/Vega ISA and latency cuts from ~118µs to ~46µs. It maps NIC RX into GPU VRAM via P2P DMA into GPU VRAM, uses a lock‑free SPSC GPU-side queue to hand per‑packet verdicts back to the NIC, and at ~32 Mpps cuts system CPU from 40% to 17–20%.
AI-written summary · about
chapters (16)
- 0:00KNOD project overview
- 1:25CPU cost in packet processing
- 3:34KNOD cost-performance
- 4:37Prototype performance and CPU savings
- 6:22KNOD programming model and XTP
- 7:48Conventional RX path limits
- 9:12Direct RX with GPU memory
- 10:17BPF to GPU JIT
- 13:32GPU shader execution
- 16:44BPF to GPU mapping
- 20:14Implementation status and plans
- 21:21Q&A: Multi-GPU support
- 23:30Q&A: Compiler reusability
- 24:33Q&A: Performance goals
- 26:40Q&A: CPU-GPU load balance
- 27:45Closing remarks
related talks
22mKernel-Native Packet Processing on AMD GPUs: XDP Offload without Userspace Runtimes
37mEfficient and Flexible Virtual Machine Networking Through eBPF
28mXDP for TSN
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets
16mElectrode: Accelerating Distributed Protocols with eBPF
18mOpen vSwitch with AF_XDP as a userspace datapath
13mRosenBridge: A Framework for Enabling Express I/O Paths Across the Virtualization...
25mShenango: Achieving High CPU Efficiency for Latency-sensitive Datacenter Workloads